On the Formal Distinction
between Literal and Figurative Language
Udo Hahn
Katja Markert
Text Understanding Lab
Freiburg University
D-79085 Freiburg, Germany
Language Technology Group
HCRC, Edinburgh University
Edinburgh EH8 9LW, U.K.
http://www.coling.uni-freiburg.de
[email protected]
The distinction between literal and gurative language (metonymies, metaphors, etc.) is often not made formally explicit, or, if
formal criteria exist, insuÆcient. This poses problems for an adequate
computational treatment of these phenomena. The basic criterion for
delineating literal from gurative speech we propose is centered around
the notion of categorization conicts that follow from the context of the
utterance. In addition, we consider the problem of granularity, which is
posed by the dependence of our approach on the underlying ontology.
Abstract.
1
Introduction
Figurative speech comes in dierent varieties (e.g., the metonymy in example
(2) and the metaphor in example (3) below), and is typically contrasted with
literal language use (e.g., example (1)) on the basis of some notion of deviance.
(1) \The man left without paying."
(2) \The ham sandwich left without paying."
(3) \The Internet is a gold mine."
Currently, two approaches prevail, which spell out this distinction. The rst
one [19] simply regards deviation from literal reference as a suÆcient condition
for gurativeness. This does not explain, however, what the formal criteria for
deviation actually are. Thus, the discrimination of literal and gurative meaning
rests on subjective ascription.
The second approach [5, 27, 24] introduces such a formal criterion, one which
depends on the notion of \violation of norms", selectional restrictions of verbs
and nouns, in particular. Each time these are violated, e.g., through type conicts, an instance of gurative speech is encountered. As a consequence, special
reasoning patterns, such as type coercion (for metonymies [24]) or analogy-based
structure mapping (for metaphors [2, 7]), are activated in order to cope with the
triggering instance such that a reasonable interpretation can be derived that
does no longer violate the underlying constraints. The proponents of this approach present a lot of supporting evidence for their methodological claims (cf.,
e.g., example (4)) but obviously fail to cover a wide range of residual phenomena (as can be seen from the example (5) that lacks any violation though being
gurative without doubt):
(4) \I read Chaucer."
(5) \I like Chaucer."
In this paper, we subscribe to the derivational paradigm, i.e., the requirement
to generate gurative speech from specications of literal language use. Thus, we
also regard gurative meaning as a deviation from literal meaning, but will not
be content to leave the description at that. Rather than formalizing the notion of
deviation with recurrence to selectional restrictions we will base our distinction
on conceptual criteria that also incorporate the inuence from the context of
an utterance. In addition, we share the assumption that the distinction between
literal and gurative speech derives from one's conceptualization and is therefore
subjective, but we will provide a formalization of this dependency on ontology.
We will also identify at least one case of ontological dependency (the problem of
granularity) where subjectiveness can be overcome by taking additional formal
criteria into account.
2
Lexical Meaning
We will base our considerations on the notion of context-independent lexical
meaning of lexemes, from which the notions of literal and gurative meaning
in context will be derived. Lexical meaning will be a function from lexemes to
categories (concepts) of an ontology.
So, let L be the set of lexemes of a given natural language and let L L be
the subset of lexemes containing nouns, full verbs and adjectives only (e.g., man
or policeman are elements of L ). We also assume an ontology composed of a set
of concept types F = fMan, Policeman, Sandwich, : : : g, a set of instances
I = fman-1, policeman-2, : : : g, and a set of relations R = fhas-part, partof, agent, : : : g (we take a settheoretical semantics for granted as is commonly
assumed in description logics [32]). The lexical meaning lex can then be dened
as a relation lex L fF [ Ig.
While we refrain from considering the linkage between lexemes and ontological entities in depth (cf., e.g., [4] or [15]), we require the relation lex to fulll
the following properties:
0
0
0
1. If lexeme 2 L is a proper name, then a unique lexeme.i 2 F [ I with
(lexeme, lexeme.i) 2 lex exists such that lexeme.i 2 I . Thus, every
proper name is linked to a single instance in the domain knowledge base.
2. If lexeme 2 L is not a proper name, then we require a concept lexeme.con
2 F to exist such that (lexeme; lexeme.con) 2 lex . In addition, no instance
lexeme.i 2 I exists such that (lexeme, lexeme.i) 2 lex .
3. For reasons of simplicity, we will now restrict lex appropriately. If lexeme
2 L is not a proper name, we require for all i 2 lex (lexeme) that i can
be referred to by lexeme in a context-independent way. So, we assume that
reference to any i via lexeme is always possible. (We cannot, e.g., relate the
lexeme fool to Man as not every man can be referenced by fool independent
of the context.) The condition of context-independence may, however, still
0
0
0
hold for several concepts that stand in a subsumption relation to each other.
When we regard the lexeme man, this condition holds for both the concepts
Man and Policeman, as all i 2 Policeman and all i 2 man can be
referenced by man. We then regard the most general concept (here, Man) as
the lexical meaning, and, in general, consider lexical meaning as a function1
lex : L 7! F [ I . By convention, we denote lex (lexeme) by lexeme.con.
0
Lexical meaning is thus considered as a context-independent function from
lexemes to categories (concepts) of an ontology. As there is no agreement on
canonical ontologies, this mapping introduces subjective conceptualizations.
Finally, we extend our denition to words w of a discourse so that their
corresponding lexeme be w:lex 2 L . We simply assume lex (w) := lex (w:lex).
We distinguish the range of that mapping by w.i for proper names and w.con
in all other cases. Hence, the lexical meaning of the word \man" in example (1)
is given by Man.2
0
3
Literal
vs.
Figurative Meaning
While in the previous section we have been dealing with the isolated lexical
meaning of a word only, we will now incorporate the context of an utterance in
which a word appears. Hence (cf. Fig. 1), we here introduce the word w with
respect to which word w is syntactically related { w is either head or modier
of w. Such a dependency relation (either a direct one or a well-dened series
of dependency relations) at the linguistic level induces a corresponding conceptual relation r 2 R at the ontological level [25]. The conceptual relation r links
the conceptual correlates, w.sf and w .sf, of w and w , respectively. Accordingly, we may now say that w StandsF or a corresponding domain entity w.sf;
alternatively, w.sf is called the (intended) meaning of w. The comparison of
w.sf with w.con or w.i lies at the heart of the decision criterion we propose
for judging whether a reading is literal or gurative. So, in the well-known example (2), \ham sandwich" (= w) StandsF or \the man who ordered the ham
sandwich" (= w.sf), which is distinct from its lexical meaning, Ham-Sandwich
(= w.con).
We may now consider some examples to distinguish several cases how w.sf
or w.consf can be related to w.con or w.i. This will also lead us to clarify the
notion of distinctiveness between the items involved. Let w.sf be an instance
from I , and let w.consf be the least general concept such that w.sf 2 w.consf .3
This assumption will be shortcut as w.sf instance-of w.consf .
0
0
0
1
2
3
0
In order to make lex a function we assume in the case of polysemy one of several
meaning alternatives to be the primary one from which the others can be derived.
In the case of homonymy, we assume the existence of dierent lexemes which can be
mapped directly to mutually exclusive concepts.
The lexical meaning of a word w must be distinguished from the concrete referent
of w in the given discourse.
The least general concept w.consf with w.sf 2 w.consf is the intersection of all
concepts C 2 F with w.sf 2 C.
Framework for Contextual Interpretation
In the most simple case, w.sf and w.con / w.i are related by an instance-of
relation. Then w.consf = w.con holds. In the utterance
Fig. 1.
(6) \A
man
left without paying."
we have w = \man" and w = \left". Furthermore, w.sf = man-1 instance-of
= w.con = w.consf . In the example (6), we note that lexical meaning
and actual meaning coincide.
If we consider all relations other than equality as deviant, we characterize
a class of phenomena that is certainly larger than the one containing gurative
speech only. Example (7)
0
Man
(7) \A policeman left without paying. The
man
lost his job."
illustrates an anaphoric relation between \the man" and \a policeman". With
this, a subsumption relation holds between w.consf (= Policeman) and w.con
(= Man), which means that w.con is either more general or equal with w.consf .
In particular, we have (policeman-2 =) w.sf 2 w.con, but not w.sf instance-of
w.con, in general (as in example (6)).
Loosening ties a bit more, we may abandon the subsumption relation between
w.consf and w.con as in example (8):
(8) \A policeman left without paying. The
fool
lost his job."
We have (policeman-2 =) w.sf 2 w.con (= Fool), but the specialization
relation between w.consf (= Policeman) and w.con (= Fool) no longer
holds. Instead, we are set back to w.sf 2 w.consf \ w.con and, therefore,
w.consf \ w.con 6= ;. We say that w.consf and w.con are compatible, as no
categorization conict arises. This also holds for the previous examples in this
section. So, the notion of categorization conict turns out to become crucial for
our distinction between literalness and gurativeness | the latter being based on
an underlying categorization conict, whereas the former is not. We summarize
these observations in the following denition:
Denition 1 (Literalness via Syntactic Constraints).
A word w in an utterance U is used according to its literal meaning, if for
every instance w.sf 2 I which w StandsF or, one of the following two conditions
hold:
w:sf = w:i
w:sf 2 w:con
Especially, w.consf
if w is a proper name
else
(1)
(2)
\ w.con 6= ; holds for non-proper nouns.
We here restrict the notion of gurative speech to those relationships between
w.sf and the lexical meaning of w in terms of w:con, which are not inclusive
ones. A literal use of the word w for an instance w.sf instance-of w.consf is
only possible, if w.consf \ w:con 6= ;. If, however, a categorization conict
occurs, i.e., w.consf \ w:con = ; holds, we call the use of w gurative. Such a
gurative use is illustrated by the word \ham sandwich" in example (2) or \gold
mine" in (3). Two consequences of this denition deserve special mention:
1. We can determine exactly the place where subjectivity comes in when a
distinction between literalness and gurativeness is made | it is mirrored by
subjectivity in categorization. \fool" in example (8) can only be considered
as literal, if the concepts Fool and Policeman are considered as being
compatible (in the settheoretic sense introduced above). If one does not share
this conceptualization, this usage of \fool" must be considered as gurative
(or even absurd). Thus, we capture the subjectivity of gurativeness formally
in the ontological premises, not via intuitive considerations.
2. It is also important to note that Denition 1 does not depend on the violation
of selectional restrictions. The example (5) (\I like Chaucer.\) allows for
the same analysis as example (4) (\I read Chaucer.\), because the intended
patient of like are, in both cases, Writings-by-Chaucer (= w.sf), although
this is not indicated by selectional restrictions at all. In both cases w.consf
\ w:con = ;, i.e., gurativeness holds.
4
Granularity
The (non-)inclusion criterion we have set up for the distinction between literal
and gurative usage of words in Denition 1 introduces a strong tie to the underlying ontology. The problem this might cause lies in granularity phenomena
of domain knowledge bases and in the general question whether every dierence
in conceptualization induces dierent literal { gurative distinctions. Given different levels of granularity, it may well happen that a word w StandsF or an
instance w.sf instance-of w.consf with w.consf \ w:con = ;, though, intuitively, one would rate the usage of w as a literal one. Let us illustrate this case
with two examples. Assume we have a knowledge base KB1 in which Cpu happens to be part-of the Motherboard, while Motherboard itself turns out
to be part-of Computer. If we analyze the example
(9) \The CPU of the computer . . . "
accordingly, we end up with the determination of a gurative usage for \computer", since Motherboard \ Computer = ; (cf. Fig. 2).
Fig. 2.
Example (9) Assuming KB1
If we assume, however, an ontological representation in a domain knowledge
base KB2 such that Cpu is an immediate part-of the Computer, then we
derive a literal usage for w (cf. Fig. 3).
Fig. 3.
Example (9) Assuming KB2
To get rid of the dependence on knowledge base granularity, to a certain
degree at least, we may derive a weaker condition of literalness from Denition
1. To achieve this, we state that w.sf and w .sf are related by a conceptual
relation r (technically, w .sf r w.sf). Thus, for literal usage of w we require:
0
0
w0 :sf r w:i if w is a proper name
9i 2 w:con: w0 :sf r i else
(3)
(4)
(3) immediately follows from (1) in Denition 1, since w .sf r w.sf (= w.i)
holds. (4) can be deduced from (2) by dening i := w.sf.
Since these conditions provide weaker conditions of literal language use than
the ones we have agreed upon in Denition 1, all literal usages determined by
the strong condition still remain literal (in particular, example (9) is considered
a literal usage of \computer" given KB2 ). Considering the granularity eects for
example (9) with respect to KB1 , however, we may determine the literal usage
of \computer" by the following consideration. As we know that Cpu is partof Motherboard, and Motherboard is part-of Computer, we conclude
0
with the transitivity of the part-of relation4 that Cpu is part-of Computer.
Hence, criterion (4) is fullled. In contradistinction to the examples discussed
previously, we do not have w.sf 2 w.con (cf. criterion (2) from Denition 1), as
w.sf = motherboard-2 holds. So by moving from the strict criteria in Denition
1 to the weaker ones stated by criteria (3) and (4) we are able to incorporate
granularity phenomena of knowledge bases. Note however that we cannot abandon completely our dependence on (the encodings of) the knowledge layer, since
we exploit knowledge about the transitivity of relations to solve the problem.
5
Figurativeness and Reference
One might argue that the problem just discussed, the dependence of the distinction between literal and gurative usage on knowledge base structures, follows
from the denition of StandsF or. Accordingly, some researchers [19] have proposed to build the denition of gurative speech upon the notion of reference.
The assumption being made is that w uniquely refers to a knowledge base item
w.ref instance-of w.conref and that gurativeness results from the deviation
of this reference from literal meaning. Although their notion of deviance is not
formalized, referentially-based literalness can now be dened straightforwardly
in our approach by proceeding exactly along the lines of Denition 1:
Denition 2 (Literalness in the Referential Approach).
A word w is called
literal in the referential approach,
w:ref = w:i
w:ref 2 w:con
if w is a proper name
else
if:
(5)
(6)
Without doubt, we here circumvent the granularity problem, since no change
in reference occurs for example (9), no matter whether KB1 or KB2 is assumed.5
But the reference approach runs into severe problems when one considers, e.g.,
classical examples of metonymies such as
(10) \I like to read
Chaucer. He
was a great writer."
We have w = \Chaucer" as a typical example for a writer-for-writings metonymy
[19]. The assumption to link literal/gurative usage to reference relations is
awed by the fact that w = \Chaucer" does not refer to the \works of Chaucer",
because in this case the referentially determined anaphor \He" could not be resolved. In particular, we have Chaucer.ref = Chaucer, therefore w.ref = w.i.
Hence, \Chaucer" must be considered, counterintuitively, as a literal use according to criterion (5) (similar problems have been discussed at length by Stallard
4
5
We are aware of empirical observations about the transitivity of part-whole relations
[3, 31], in particular those pertaining to the claim that any of the subrelations of partwhole are transitive, while the general part-whole relation usually is not. When we talk
about the transitivity of part-whole, we mean this constrained type of transitivity.
Note that this denition is, nevertheless, still dependent on the knowledge base and
on the lexical meaning of w.
[27]). Given our context-dependent denitions (1 or 3), we get w.sf = Writingsby-Chaucer so that w.sf 6= Chaucer. Thus, w = \Chaucer" is analyzed gura-
tively in our approach.6
Summarizing, we combine criteria (1) to (6) by the following conditions:
1. A word w is used in its literal meaning in all denitions, if w.sf 2 w.con
(analogously, w.sf = w.i) for all w.sf, and w.ref 2 w.con hold (combining
the referential Denition 2 and Denition 1 with respect to literal usage).
2. If w.sf 62 w.con (analogously, w.sf 6= w.i) for some w.sf (with respect to
a relation r and another word w ), but w.ref 2 w:con (w.ref = w.i), two
cases must be distinguished:
{ In cases of granularity eects criterion (4) holds. By this, an i 2 w.con
exists with w .sf r i (analogously, w .sf r w.i). We can include this in
our denition of literal usage as its analysis is only due to implications
a particular ontology design brings to bear.
{ In cases of gurative speech like the one in example (10) the criteria
(4) / (3) do not hold. We may include these cases into our denition of
gurative usage:
3. A word w is used in its gurative meaning according to the syntactic and
the referential denition, if w.ref 62 w.con holds and there exists a w.sf 62
w.con. This is the case, e.g., in example (2).
So far, we have only considered the gurative usage of a word w. We may
end up by dening U as a gurative utterance, if it contains at least one word w
which is used in a gurative way.
0
0
6
0
Putting Theory to Practice { Metonymy Resolution
So far, we considered the gurativeness from an entirely theoretical perspective.
Our interest in this topic can be traced back, however, to practical requirements
which emerge from SynDiKATe, a text understanding system for processing
product reviews from the IT domain as well as medical nding reports [10].
The need to deal with a particular type of gurative language, metonymic
expressions, becomes evident when one considers the quantitative distribution
of referential text phenomena we determined for a random sample of 26 texts.
In 606 utterances we encountered 103 metonymies together with 291 nominal
anaphora and 351 bridging anaphora [20]. With a metonymic expression encountered in almost every sixth utterance, an uncontroversial need for dealing with
this problem is demonstrated. We also collected empirical evidence that a 16%
increase in the accuracy of the anaphora resolution procedure can be achieved
by the incorporation of a metonymy resolution process (due to the overlapping
of metonymies and nominal anaphora).
SynDiKATe builds on a fully lexicalized dependency grammar [11]. Parse
trees link lexical nodes via a well-dened set of dependency relations (cf. genitiveattribute as illustrated in Figures 2 and 3). For semantic interpretation [25], the
6
A methodology for intersentential anaphora resolution, given these interpretations,
is described in [28].
conceptual correlates of the lexical items in the dependency graph need to be
linked via conceptual relations. This linkage is based on conceptual constraints.
One major source of these constraints are dependency relations, since they are
often characterized by a hard-wired mapping onto general conceptual relations.
For instance, the dependency relations subject or direct-object, by convention,
map to the conceptual relations agent and patient or patient, respectively. There
are types of dependency relations, e.g., genitive-attribute, which have no associated conceptual constraints. In such cases, a fairly unconstrained search of the
knowledge base for proper linkage is performed resulting, e.g., in the determination of the part-of relation (chain) between Cpu-1 and Computer-2 in Fig. 3 for
the lexical correspondents \Cpu" and \computer" linked via genitive-attribute.
We compute possible StandsFor relationships relevant for the distinction between literal and gurative language by looking at the paths found as a result
of the search of the knowledge base for conceptual relations. Especially, we look
at the intermediate concept nodes a path contains in order to determine the
compatibility between those concept and the lexical meaning. For illustration
purposes, a few examples of paths will be discussed subsequently. When the
dierent interpretations are computed and distinguished, these distinctions can
then be used for disambiguation and preferential ranking.7
Consider the following two examples already introduced in this paper.
(11) \The ham sandwich left (without paying)."
(12) \(I) like Chaucer."
In example (11) we concentrate on the subject dependency relation between
\ham sandwich" and \left". As already remarked, subject maps conceptually
to agent (for active voice) or patient (for passive voice) roles. Concentrating
on the active voice case, the concept Leave has associated a conceptual role,
leave-agent, a subrole of agent, which has a range constraint on Persons. A
Person may order a Product, a possible product being eatables such as
Ham-Sandwiches. A possible linkage between the conceptual correlates of \ham
sandwich" and \left" is, therefore, given by the following path: p11 = (Leave {
leave-agent { Person { orders { Ham-Sandwich).
In example (12) we concentrate on the patient relation between \like" and
\Chaucer" as induced by the direct-object relation. In this case, the concept Like
is associated with the conceptual role, like-patient, a subrole of patient, which has
no range constraint other than the most general concept AnyThing. AnyThing
subsumes Persons, which Chaucer is an instance of, so that the path p112 =
(Like { like-patient { Chaucer) just expresses that Chaucer as a person is liked.
In addition, Anything also subsumes Books, which are written-by a particular
Author, one of those being Chaucer. In this case, we have constructed another
path: p212 = (Like { like-patient { Book { written-by { Chaucer).
So when possible paths between w.con and w0.con have been determined the
proposed gurativeness criterion can be used to determine whether the readings
7
Of course, we restrict the search through the knowledge base, e.g., by conceptual
connectivity criteria (cf. [20] for details).
that can be derived from the path are gurative or literal. Let us look at our
example path p11 = (Leave { leave-agent { Person { orders { Ham-Sandwich)
again. Possibilities for the StandsF or items left.sf and ham-sandwich.sf and
the conceptual relation r corresponding to the syntactic subject relation are
{ left.sf 2 Leave (= Left.con), r = leave-agent and sandwich.sf 2 Person
(with Person \ Ham-Sandwich = ;).
{ left.sf 2 Person (with Person \ Leave = ;), r = orders and sandwich.sf
2 Sandwich (= Sandwich.con).
The rst possibility gives rise to a literal reading of left and a gurative
reading of ham sandwich when compared to our criterion whereas the second
gives rise to a gurative reading of left and a literal reading of ham sandwich.
In the Chaucer example possibilities for the StandsF or items like.sf and
chaucer.sf and the conceptual relation r corresponding to the syntactic directobject relation include
{ like.sf 2 Like (= like.con), r = like-patient and chaucer.sf = Chaucer
(= Chaucer.i). This is the only reading that can arise by taking the path
p112 into account as there are only two concepts and one relation in the path.
The next two readings arise when looking at p212 :
{ like.sf 2 Like (= like.con), r = like-patient and chaucer.sf 2 Book
(with Book and Chaucer.i being incompatible).
{ like.sf 2 Book (with Book \ Like = ;), r = written-by and chaucer.sf
= Chaucer (= Chaucer.i).
The rst possibility accounts for a literal reading of Chaucer and like, whereas
the second and third account for a gurative reading for Chaucer or like.
It is important to note that the disambiguation of dierent readings may
proceed along several lines such that this process is not aected by the criterion of gurativeness itself. The determination of possible readings can, e.g., be
combined with all sorts of disambiguation criteria or search heuristics such as
1. Preference of literal readings over gurative ones, thus preferring the rst
reading in the Chaucer example over the other two readings. One can show
that an approach based on this criterion only amounts to an approach considering selectional restriction violations as necessary for establishing gurative
readings.
2. Additional world knowledge that determines that a literal reading in the
Chaucer example is invalid when uttered in 1999.
3. Knowledge about anaphora and discourse relations might be helpful for the
Ham-Sandwich example, e.g., such that a record of discourse entities is
kept from which one may derive that a person has ordered a sandwich who
has already been introduced into the discourse. So that person is likely to be
referred to again, thus making the gurative reading of ham sandwich more
likely than the gurative reading of left.
4. Statistical knowledge about linguistic structures that make it more likely
that a noun phrase be used metonymically than a verb phrase, thus preferring
the second reading over the third reading in the Chaucer example.
All these criteria can be combined with our gurativeness criterion so that the
criterion can also provide a testbed for preference heuristics and search restrictions when analyzing gurative language. In the framework of SynDiKATe,
we have, rst, developed path patterns inferred from the gurativeness criterion that make the distinction between gurativeness and literalness based on
paths between knowledge base concepts easier to determine. A simple example would be that paths of length 1 always mirror literal readings of both lexical
items involved.8 We then developed various search and disambiguation heuristics
which take into account the following features (for details, cf. [20]):
{ Common (or, typical) gurative relations are preferred over not so common
ones, thus preferring paths including patterned relations.
{ Discourse criteria preferring readings allowing for anaphoric readings over
ones that do not.
{ Only when those criteria do not apply, we prefer literal readings over gurative ones.
It is important to note that these kinds of comparisons can only be achieved
when a well-founded distinction of gurative and literal language is possible,
which should not rely on the heuristics themselves, e.g., restricting gurative
speech to cases where a sortal conict occurs. Instead, an independent ontological
criterion is necessary that can be combined freely with reasonable heuristics.
7
Related Work
We consider as the main contribution of this paper the introduction of a formal
notion of deviance that is general and simple. To the best of our knowledge, no
comparable work has been done so far on this issue. Although there exist formal
characterizations of metaphors [13, 7] in the framework of analogical reasoning,
these are entirely self-contained, i.e., they account for structural properties of
metaphors (e.g., constraints on domain mappings, aptness conditions), rather
than dealing with the distinction between literal and gurative speech. Even
more serious from a natural language understanding point of view, concrete
metaphorical utterances cannot be translated into the formal systems provided
for metaphor explanation due to the lack of concrete mappings between the
lexical and conceptual level [13].
The state of the art aiming at a clarication of what constitutes, e.g., a
metonymy is characterized by a quote from Lako and Johnson [19]. They nd
that it is characterized by the use of \one entity to refer to another that is related
to it". It is not clear at all what kind of entities they are talking about. One must
8
There are other literal paths as well which account for granularity problems.
assume that they are referring to some kind of signs because it is not obvious
how other entities can refer to something. But even if we restrict the entities
to linguistic signs there is no restriction on the kind of relatedness between the
objects. For example, relatedness might include class inclusion, similarity or partwhole relations, but only the latter are included in metonymy in general and the
examples Lako and Johnson put forward suggest that it is this conventional
kind of metonymy they are talking about. The same shadowy denitions of
gurative language are then often adopted by theoretical linguists [29, 17, 30,
23], as well as computational linguists [5, 21]. This leads to the fact that it is
mostly not clear at all, which phenomena are treated by these approaches and
how they discriminate dierent varieties of gurative speech from others.
In addition, a tendency can be observed in more formal approaches { pressed
by the need to nd computationally feasible denitions of metaphor or metonymy
{ to consider gurative language a violation of selectional restrictions [2, 5, 12,
24] or communicative norms [9, 26]. Such an approach equates an often used
triggering condition, viz. constraint violation, with the phenomenon of gurative
language (or, subsets, like metonymies). So, it confuses the possible, but not
necessary eects of a phenomenon with the phenomenon to be explained.
Despite the lack of formal rigor in previous work, it is worth to investigate
how our formal criterion is compatible with other views on gurative speech from
cognitive linguistics in general. The tendency to see gurative speech rooted in
conceptual categories, as we do, is becoming consensus. The main trend is, e.g.,
to treat metaphors as a means of categorization by way of similarity [8] and,
in principle, to retrace gurative speech to cognitive procedures involving categorization and (subjective) experience [19, 6, 18]. So, Lako and Johnson see
metaphors rooted in our way of conceptualization via mappings. Kittay [17]
and Turner [30] regard some kind of conceptual incompatibilities as the basis of
metaphorization. Nevertheless, they do not explicate their theory of categorization and incompatibility nor do they recognize that these incompatibilities are
relevant for other kinds of gurative speech, as well as for metaphors in the strict
sense of the word. The dependence of lexical, literal and gurative meaning on
ontologies is, therefore, realized, but no explicit formal treatment is given of particular problems resulting from the ontological presuppositions. We here regard
the problem of granularity arising form ontological dependence and subjective
experience and propose a solution by relaxing criteria for literal meaning.
The second major contribution of the paper derives from the formal status
of our distinction between literal and gurative meaning. Once such a formal
notion is given, it allows us to characterize subjectivity, so far an entirely informal notion, by reference to the particular ontology underlying the language
understanding process. The current perspective on the ontological basis is rather
static, viz. dierent ontologies account for cases in which one person considers
an utterance to be literal, while another considers it to be gurative. Our proposal introduces exibility into such an assessment. We aim at adapting dierent
ontologies such that by way of abstracting away dierent granularities of representation structures (e.g., by generalizing more ne-grained representations to
a coarser grain size, as in criterion (4)) disagreement might turn into consensus
(e.g., considering example (9)). Contrary to that, the majority of researchers in
our eld of study attribute the dierence in opinion to the existence of dierent, incompatible ontologies, and leave it with that explanation without further
attempt at smoothing [19, 30].
The third major proposal we make relates to the contextual embedding of
gurative speech. The criterion we formulate is based on syntactic relations only
that guide conceptual interpretation. In particular, and unlike most algorithmic
accounts [22, 5, 12, 24], it does not rely at all upon the violation of selectional restrictions (for a notable exception, cf. [21]), since this criterion accounts only for
a subset of the phenomena naturally recognized as gurative language. In addition, the syntax-based proposal we make avoids to consider reference changes as
an indicator of gurativeness as is commonly assumed [19]. It is inspired by Fauconnier's [6] \connector" function. Though this proposal aims at an embedding
of gurative language into syntax, there exists no formalization of this notion
in relation to an established grammar framework as we do nor is the notion of
conceptual incompatibility formalized. A more formal criticism of the reference
changes proposal was made by Stallard [27] who, nevertheless, then only dealt
with gurative language, violating sortal constraints.
Our approach is also compatible with viewing gurative language as regular
and not violating linguistic norms. Whereas literal language is grounded in inclusion relations to lexical or context-independent meaning, gurative language
is grounded in other relations to lexical meaning. These can, nonetheless, be
systematic and conventionalized relations like part-whole relations or obvious
similarities. Although we have not spelled out these relations, there is no need
to claim that inclusion relations are prior or preferred to other relations. Literal as well as gurative speech are both grounded in structured relations in
categorization. This is in accordance with the conventional metaphor view rst
stipulated by Lako and Johnson. It is also in accordance with psycholinguistic
research showing that gurative speech is in most cases as easily understood as
literal speech. This is especially the case if the instance of gurative speech is
conventional, i.e. grounded in conventional, systematic and pervasive ontological
relationships [1]. The essence of this is that pervasive and structured relations
or relations made salient by the context [14] may be as easily available to comprehension as inclusion relations.
8
Conclusion
In this paper, we have drawn a distinction between literal and gurative speech
which is based on formal criteria. These are grounded in the solid framework
of description logics, in particular, by relying on its settheoretical semantics.
We aimed at a comprehensive distinction, one that covers the whole range of
word-centered gurative language use (i.e., metonymies, metaphors, irony) using
general and simple criteria. A major determinant for gurative interpretation is
given by the context of an utterance. As a consequence, contextual criteria are
at the core of any adequate account of gurativeness. Our contextual criteria
are independent of the notion of selectional restrictions, but dependent on the
conceptual interpretation of syntactic relations, in general. Another crucial condition of whether language use is considered literal or gurative is introduced
by the particular ontology referred to. While earlier formalist approaches appeal
to semantic types, sortal constraints, etc., this is not fully convincing, since the
entire structure and granularity of the theory of the domain being talked about
contributes to the understanding process, whether literally or guratively based.
In particular, we captured the notion of subjectivity in ontological premises and
explained how granularity problems may be overcome.
The model we presented does currently not account for neologisms, as those
have no a priori lexical meaning, and many tricky cases of quantication and
the use of proper names. From a more technical perspective, we have also not
scrutinized the dierent kinds of relations that are still required to hold between
w.consf and w.con, if w.consf \ w.con = ;. So, a necessary condition for gurative speech has been established that needs to be supplemented by suÆcient
ones. We also have no criteria available right now to distinguish between various
types of gurative speech (e.g., metaphors vs. irony). By this, we mean specications concerning similarity relations for metaphors, contrast relations for irony,
and contiguity relations for metonymies (cf., e.g., [16]). Finally, we stop short
of distinguishing between innovative gurative speech (like in the ham sandwich
example) and conventionalized gurative speech (systematic polysemy [24, 23]).
Acknowledgements. K. Markert was a member of the Graduate Program Human and
Machine Intelligence at Freiburg University, funded by DFG.
References
1. Dawn G. Blasko and Cynthia M. Connine. Eects of familiarity and aptness on
metaphor processing. Journal of Experimental Psychology: Learning, Memory and
Cognition, 19(2):295{308, 1993.
2. Jaime G. Carbonell. Metaphor: an inescapable phenomenon in natural-language
comprehension. In W. G. Lehnert and M. H. Ringle, editors, Strategies for Natural
Language Processing, pages 415{434. Hillsdale, N.J.: Lawrence Erlbaum, 1982.
3. Roger ChaÆn. The concept of a semantic relation. In A. Lehrer and E. F. Kittay,
editors, Frames, Fields and Contrasts, pages 253{288. Hillsdale: Erlbaum, 1992.
4. D. A. Cruse. Lexical Semantics. Cambridge: Cambridge University Press, 1986.
5. Dan C. Fass. met : a method for discriminating metonymy and metaphor by
computer. Computational Linguistics, 17(1):49{90, 1991.
6. Gilles Fauconnier. Espaces Mentaux. Paris: Les Editions de Minuit, 1984.
7. Dedre Gentner, Brian Falkenhainer, and Janice Skorstad. Viewing metaphors as
analogy: the good, the bad, and the ugly. In Y. Wilks, editor, Theoretical Issues in
Natural Language Processing, pages 171{177. Hillsdale: Lawrence Erlbaum, 1989.
8. Raymond Gibbs. Categorisation and metaphor understanding. Psychological Review, 99(3):572{577, 1992.
9. Herbert P. Grice. Logic and conversation. In P. Cole and J. Morgan, editors,
Syntax and Semantics, volume 3, pages 41{58. New York: Academic Press, 1975.
10. Udo Hahn and Martin Romacker. Text understanding for knowledge base generation in the SynDiKATe system. In Database and Expert Systems Applications.
Proceedings of the 10th International Conference { DEXA'99. Florence, Italy, August 30 - September 3, 1999. Berlin: Springer, 1999.
11. Udo Hahn, Susanne Schacht, and Norbert Broker. Concurrent, object-oriented
natural language parsing: the ParseTalk model. International Journal of HumanComputer Studies, 41(1/2):179{222, 1994.
12. Jerry R. Hobbs, Mark E. Stickel, Douglas E. Appelt, and Paul Martin. Interpretation as abduction. Articial Intelligence, 63(1/2):69{142, 1993.
13. Bipin Indurkhya. Constrained semantic transference: a formal theory of metaphors.
In A. Prieditis, editor, Analogica, pages 129{157. Morgan Kaufmann, 1988.
14. A. Inho, S.D. Lima, and P.J. Carroll. Contextual eects on metaphor comprehension in reading. Memory and Cognition, 12:558{567, 1984.
15. Ray Jackendo. Semantic Structures. Cambridge, MA: MIT Press, 1990.
16. Roman Jakobson and Morris Halle. Fundamentals of Language. The Hague, NL:
Mouton Publishers, 1956.
17. Eva F. Kittay. Metaphor: Its Cognitive Force and Linguistic Structure. Oxford,
U.K.: Clarendon Press, 1987.
18. George Lako. Women, Fire, and Dangerous Things. What Categories Reveal
about the Mind. Chicago, IL: University of Chicago Press, 1987.
19. George Lako and M. Johnson. Metaphors We Live By. Chicago, IL: University
of Chicago Press, 1980.
20. Katja Markert and Udo Hahn. On the interaction of metonymies and anaphora.
In IJCAI'97 { Proceedings of the 15th International Joint Conference on Articial
Intelligence, pages 1010{1015. Nagoya, Japan, August 23-29, 1997.
21. James Martin. Computer understanding of conventional metaphoric language.
Cognitive Science, 16:233{270, 1992.
22. Peter Norvig. Marker passing as a weak method for inferencing. Cognitive Science,
13(4):569{620, 1989.
23. Georey Nunberg. Transfers of meaning. Journal of Semantics, 12:109{132, 1995.
24. James Pustejovsky. The Generative Lexicon. Cambridge, MA: MIT Press, 1995.
25. Martin Romacker, Katja Markert, and Udo Hahn. Lean semantic interpretation.
In IJCAI'99 { Proceedings of the 16th International Joint Conference on Articial
Intelligence. Stockholm, Sweden, July 31 - August 6, 1999.
26. John R. Searle. Metaphor. In Andrew Ortony, editor, Metaphor and thought, pages
93{123. Cambridge, U.K.: Cambridge University Press, 1979.
27. David Stallard. Two kinds of metonymy. In Proceedings of the 31st Annual Meeting
of the Association for Computational Linguistics, pages 87{94. Columbus, Ohio,
USA, 22-26 June, 1993.
28. Michael Strube and Udo Hahn. Functional centering. In ACL'96 { Proceedings of
the 34th Annual Meeting of the Association for Computational Linguistics, pages
270{277. University of California at Santa Cruz, California, 24-27 June, 1996.
29. Roger Tourangeau and Robert Sternberg. Understanding and appreciating
metaphors. Cognition, 11:203{244, 1982.
30. M. Turner. Categories and analogies. In D. H. Helman, editor, Analogical Reasoning, pages 3{24. Dordrecht, NL: Kluwer Academic Publishers, 1988.
31. M. Winston, R. ChaÆn, and D. Herrmann. A taxonomy of part-whole relationships. Cognitive Science, 11:417{444, 1987.
32. William A. Woods and James G. Schmolze. The Kl-One family. Computers &
Mathematics with Applications, 23(2/5):133{177, 1992.